Approximate Nearest Neighbor Search Indexes [experimental]

The main task that indexes achieve is to quickly find nearest neighbors for multidimensional data. An example of such a problem can be finding similar pictures (texts) for a given picture (text). That problem can be reduced to finding the nearest embeddings. They can be created from data using UDF.

The next queries find the closest neighbors in N-dimensional space using the L2 (Euclidean) distance:

SELECT * 
FROM table_name 
WHERE L2Distance(Column, Point) < MaxDistance 
LIMIT N

SELECT * 
FROM table_name 
ORDER BY L2Distance(Column, Point)
LIMIT N

But it will take some time for execution because of the long calculation of the distance between TargetEmbedding and all other vectors. This is where ANN indexes can help. They store a compact approximation of the search space (e.g. using clustering, search trees, etc.) and are able to compute approximate neighbors quickly.

Indexes Structure

Approximate Nearest Neighbor Search Indexes (ANNIndexes) are similar to skip indexes. They are constructed by some granules and determine which of them should be skipped. Compared to skip indices, ANN indices use their results not only to skip some group of granules, but also to select particular granules from a set of granules.

ANNIndexes are designed to speed up two types of queries:

Type 1: Where

SELECT * 
FROM table_name 
WHERE DistanceFunction(Column, Point) < MaxDistance 
LIMIT N

Type 2: Order by

SELECT * 
FROM table_name [WHERE ...] 
ORDER BY DistanceFunction(Column, Point) 
LIMIT N

In these queries, DistanceFunction is selected from distance functions. Point is a known vector (something like (0.1, 0.1, ... )). To avoid writing large vectors, use client parameters. Value - a float value that will bound the neighbourhood.

:::note ANN index can't speed up query that satisfies both types (where + order by, only one of them). All queries must have the limit, as algorithms are used to find nearest neighbors and need a specific number of them. :::

:::note Indexes are applied only to queries with a limit less than the max_limit_for_ann_queries setting. This helps to avoid memory overflows in queries with a large limit. max_limit_for_ann_queries setting can be changed if you know you can provide enough memory. The default value is 1000000. :::

Both types of queries are handled the same way. The indexes get n neighbors (where n is taken from the LIMIT clause) and work with them. In ORDER BY query they remember the numbers of all parts of the granule that have at least one of neighbor. In WHERE query they remember only those parts that satisfy the requirements.

Create table with ANNIndex

This feature is disabled by default. To enable it, set allow_experimental_annoy_index to 1. Also, this feature is disabled on ARM, due to likely problems with the algorithm.

CREATE TABLE t
(
  `id` Int64,
  `data` Tuple(Float32, Float32, Float32),
  INDEX ann_index_name data TYPE ann_index_type(ann_index_parameters) GRANULARITY N
)
ENGINE = MergeTree
ORDER BY id;

CREATE TABLE t
(
  `id` Int64,
  `data` Array(Float32),
  INDEX ann_index_name data TYPE ann_index_type(ann_index_parameters) GRANULARITY N
)
ENGINE = MergeTree
ORDER BY id;

With greater GRANULARITY indexes remember the data structure better. The GRANULARITY indicates how many granules will be used to construct the index. The more data is provided for the index, the more of it can be handled by one index and the more chances that with the right hyper parameters the index will remember the data structure better. But some indexes can't be built if they don't have enough data, so this granule will always participate in the query. For more information, see the description of indexes.

As the indexes are built only during insertions into table, INSERT and OPTIMIZE queries are slower than for ordinary table. At this stage indexes remember all the information about the given data. ANNIndexes should be used if you have immutable or rarely changed data and many read requests.

You can create your table with index which uses certain algorithm. Now only indices based on the following algorithms are supported:

Index list

Annoy

Annoy

Implementation of the algorithm was taken from this repository.

Short description of the algorithm: The algorithm recursively divides in half all space by random linear surfaces (lines in 2D, planes in 3D etc.). Thus it makes tree of polyhedrons and points that they contains. Repeating the operation several times for greater accuracy it creates a forest. To find K Nearest Neighbours it goes down through the trees and fills the buffer of closest points using the priority queue of polyhedrons. Next, it sorts buffer and return the nearest K points.

Examples:

CREATE TABLE t
(
  id Int64,
  data Tuple(Float32, Float32, Float32),
  INDEX ann_index_name data TYPE annoy(NumTrees, DistanceName) GRANULARITY N
)
ENGINE = MergeTree
ORDER BY id;

CREATE TABLE t
(
  id Int64,
  data Array(Float32),
  INDEX ann_index_name data TYPE annoy(NumTrees, DistanceName) GRANULARITY N
)
ENGINE = MergeTree
ORDER BY id;

:::note Table with array field will work faster, but all arrays must have same length. Use CONSTRAINT to avoid errors. For example, CONSTRAINT constraint_name_1 CHECK length(data) = 256. :::

Parameter NumTrees is the number of trees which the algorithm will create. The bigger it is, the slower (approximately linear) it works (in both CREATE and SELECT requests), but the better accuracy you get (adjusted for randomness). By default it is set to 100. Parameter DistanceName is name of distance function. By default it is set to L2Distance. It can be set without changing first parameter, for example

CREATE TABLE t
(
  id Int64,
  data Array(Float32),
  INDEX ann_index_name data TYPE annoy('cosineDistance') GRANULARITY N
)
ENGINE = MergeTree
ORDER BY id;

Annoy supports L2Distance and cosineDistance.

Setting search_k (default LIMIT * NumTrees) determines how many nodes the Annoy index will inspect during SELECT queries. The setting can be used to balance performance and accuracy at runtime.

Example:

SELECT * 
FROM table_name [WHERE ...] 
ORDER BY L2Distance(Column, Point) 
LIMIT N
SETTING ann_index_select_query_params=`k_search=100`

6.9 KiB Raw Blame History